Cost per task, by agent and skill
Know what every agent task costs.
Every task your agents run leaves a bill: tokens, time, retries. RChursin_ itemises it by agent and skill, so you can see what pays for itself and what doesn’t.
Why idle skills cost something
Fig. 1 · Tokens per skill
~100
What one skill’s name and description add to every turn, used or not. Anthropic docs
Fig. 2 · Tokens per turn
~8,000
Loaded before the first word of work when a repo has 80 skills.
Fig. 3 · Agents in use
3 of 12
Agents that did the work in our founder’s own setup of twelve.
The ledger
One line for every agent and skill.
Runs, success rate and cost per finished task, with a verdict you can act on.
| Name | Runs | Success | Cost / task | Verdict | Note |
|---|---|---|---|---|---|
| pr-review | 212 | 91% | $0.92 | keep | Used daily. Finishes the job. |
| test-fixeragent | 34 | 41% | $3.20 | fix | Fails more than half the time. |
| api-docs | 18 | 83% | $0.61 | merge | Does the same job as openapi-writer. |
| release-notes | 48 | 88% | $0.40 | keep | Cheap and reliable. |
| db-migration-helper | 0 | — | — | cut | No runs in 45 days. |
| ui-snapshot | 3 | 67% | $1.10 | fix | Only runs when typed as /ui-snapshot. |
| legacy-refactoragent | 0 | — | — | cut | No runs in 60 days. |
Fig. 4 · Cost per finished task
One agent costs over three times the median.
test-fixer costs $3.20 a finished task, 3.5× the median, and still fails more often than it works. That’s the first one to fix.
- main session$1.12
- pr-review$0.92
- code-explorer$0.74
- test-fixer$3.20
- docs-writer$0.58
How it works
From session to receipt in three steps.
- 01
Connect
Turn on your agent’s telemetry, or point us at your local session files. Prompts stay out by default.
- 02
Map
We match each session to the agent and skill files in your repo, and to the result: task done, tests passing, PR merged.
- 03
Act
A weekly report says what to keep, cut, merge or fix, with pull requests ready to review.
“I had 12 agents and 80 skills. I used 3 agents and about 10 skills. I couldn’t tell which of the rest were dead weight, so I built RChursin_.”
FAQ
About the numbers.
Claude Code already shows usage. Why do I need this?
Built-in tools show one machine and the last few days. RChursin_ shows your whole team, keeps the history, and ties every agent and skill to results, not just to how often it ran.
Does my code or my prompts leave my machine?
The local scan runs on your machine and sends nothing. Team mode receives usage counts, costs and names. Claude Code keeps prompts out of its telemetry unless you turn them on.
How do you decide something is useful?
We compare sessions that used an agent or skill with similar sessions that didn’t: did the task finish, how long did it take, and what did it cost.
Which coding agents do you support?
[CONFIRM BEFORE LAUNCH] Claude Code first. Cursor, OpenCode and Codex next.
Why does an unused skill cost anything?
Its name and description are loaded on every turn so the agent knows it exists. That’s about 100 tokens a skill. Small alone, but 80 skills is around 8,000 tokens before any work starts.
Your first receipt